LoginSign Up
← Back to all sparks
O

Ollama

AI-ASSISTANTS
Velocity6.3

Run and manage large language models locally on your own machine.

ollama.com ↗

Ollama turns into a launcher for agentic coding tools between llama.cpp and MLX upkeep

local-inferenceagentic-codingllama-cppmlxgpu-offloadmodel-support
◆Current state
Ollama's recent releases split between routine engine maintenance and a quieter, more interesting move: becoming the local runtime that installs and manages agentic coding tools. Stable builds now auto-install Claude Code and opencode, detect Codex model drift, and add thinking-capability detection, alongside continuous llama.cpp and MLX updates and GPU-offload tuning. Most of the newest activity is release-candidate churn rather than user-facing change.
◆Where it's heading
The engine work — MLX on Apple Silicon, iGPU projector offload, speculative decoding — keeps broadening hardware reach, but the 'launch' subsystem is the directional bet: Ollama positioning itself as the local backend and manager for coding agents. If that continues, Ollama becomes less a model runner and more the control point between local models and agentic dev tools.
◆Prediction
Expect the 0.31.2 line to stabilize out of release candidates soon, and further 'launch' integrations wiring additional agent front-ends to local Ollama models.

◆Recent moves

  1. 1d ago

    v0.31.2-rc2: llm: allow iGPU mmproj offload with fit padding (#16996)

    A release candidate letting integrated GPUs like GB10 and Strix Halo offload the multimodal projector via memory-fit padding, instead of forcing CLIP onto CPU. Narrow hardware enablement, still pre-release.

    View source ↗
  2. 2d ago

    v0.31.2-rc1: create: harden GGUF create flows (#17062)

    A release candidate hardening the GGUF model-creation path, plus lint. Internal robustness work with no user-facing change.

    View source ↗
  3. 2d ago

    MLX dependency update (0.31.2-rc0)

    A release candidate bumping the MLX dependency. Routine upkeep of the Apple Silicon engine.

    View source ↗
  4. 10d ago

    Tool-call JSON parsing fix; llama.cpp and MLX bumps

    The 0.30.12 stable release fixes tool-call end detection so braces inside JSON strings no longer break parsing, and refreshes llama.cpp and MLX. A real correctness gain for tool-calling workloads.

    View source ↗
  5. 13d ago

    Add Ornith 9B renderer and parser (rc)

    A release candidate adding a renderer and parser for the Ornith 9B model. Incremental model support, still pre-release.

    View source ↗
  6. 14d ago

    Auto-install Claude Code and opencode; Windows GPU fixes

    ⚡ SPARK

    The 0.30.11 stable release adds a 'launch' subsystem that auto-installs Claude Code and opencode, detects Codex model drift, and adds thinking-capability detection — alongside Windows Vulkan GPU fixes and speculative-decoding tuning. The clearest step toward Ollama managing agentic coding tools, not just serving models.

    View source ↗